Anthropic secures 2.5 GW of Nvidia Vera Rubin GPU capacity amid tight market supplies
Nvidia disclosed last Sunday that Anthropic has contracted $180 billion in infrastructure value across cloud providers and neocloud operators, including 2.5 gigawatts of Nvidia-specific AI capacity scheduled for delivery through 2028. The figure, revealed during Nvidia's September 28 Non-Deal Roadshow investor briefing, represents the first public aggregation of Anthropic's sprawling infrastructure commitments into a single contracted-value number — and the 2.5 GW Nvidia-specific slice is the metric that matters most for understanding Nvidia's own revenue trajectory.
The reason: every gigawatt Anthropic has committed to Nvidia is worth more to the chipmaker than the gigawatt before it. Nvidia said its revenue opportunity per gigawatt, based on its reference designs, has risen with each generation from roughly $18 billion during the Hopper generation to $25 billion with Grace Blackwell and now $40 billion with the forthcoming Vera Rubin platform. That progression is not primarily a GPU pricing story. It is a product-mix story — and understanding what drives it is the key to reading both Nvidia's growth and Anthropic's infrastructure strategy.
**How Nvidia Extracts More Revenue From Every Watt**
In the Hopper era, a customer buying a gigawatt of AI infrastructure was buying mostly one thing: H100 GPU units plus standard InfiniBand networking. Revenue per gigawatt was substantial — roughly $18 billion — but it was concentrated in GPU average selling price.
The Grace Blackwell transition changed the product mix. By integrating its own Arm-based Grace CPU directly with the Blackwell GPU in the GB200 and GB300 rack-scale systems, Nvidia captured CPU revenue that had previously gone to Intel and AMD. NVLink scale-up interconnect — a proprietary Nvidia technology for connecting GPUs at extreme bandwidth within a rack — became a required component of high-performance Blackwell deployments, adding networking revenue on top of GPU and CPU revenue. The Spectrum-X Ethernet and Quantum InfiniBand platforms scaled further. The result: $25 billion per gigawatt.
Rubin's path to $40 billion per gigawatt adds two more product categories. The first is a dedicated LPU (Low-Latency Processing Unit) rack — a new hardware category Nvidia is shipping specifically to handle agentic AI workloads, which require high-throughput, low-latency inference at volumes that general GPU clusters handle inefficiently. The second is an expanded software layer: CUDA, Nvidia's AI Foundry services, and Nemotron open model weights, all of which generate revenue that was not part of the Hopper bundle. SemiAnalysis, the semiconductor research firm, estimated in September 2026 that the Rubin NVL72 generates approximately 39% more annual revenue and 42% more profit per gigawatt than the strongest Grace Blackwell GB300 configuration, crediting integrated networking and interconnect architecture.
Power efficiency reinforces the economics. Nvidia has said Vera Rubin delivers 30 times higher throughput per megawatt and 35 times lower token cost per million tokens than Grace Blackwell Ultra. That means operators deploying Rubin can serve dramatically more AI inference workloads from the same power envelope — which is increasingly the binding constraint on AI infrastructure expansion as power grids strain under data center electricity demand.
There is, however, one condition the $40 billion per gigawatt figure depends on: customers must buy the full stack. Nvidia's revenue opportunity per gigawatt is calculated based on its reference designs — meaning a deployment that includes GPU, CPU, LPU, NVLink, Spectrum-X networking, and software. Customers who purchase only Rubin GPUs without the accompanying CPU, LPU, and networking layers will not generate $40 billion per gigawatt for Nvidia. The rate at which customers adopt the full-stack DSX model — rather than cherry-picking GPUs alone — is the single most important variable in whether Nvidia's per-gigawatt revenue target materializes in practice.
**Anthropic's Infrastructure Portfolio: From AWS to Australia**
Nvidia's disclosure of Anthropic's $180 billion contracted value appears to aggregate infrastructure commitments accumulated across 2025 and 2026. The largest individual deal is a more than $100 billion, ten-year commitment with AWS announced April 20, 2026, which secures up to 5 gigawatts of capacity for training and deploying Claude, including more than one million Trainium2 chips.
A separate deal announced around September 22, 2026 with UK-based neocloud operator Nscale covers approximately $45 billion over six years for around 460 megawatts of capacity at a West Virginia data center facility, expected to come online by late 2027 and powered by Nvidia's Vera Rubin-generation systems. In early September 2026, reporting emerged of a $35 billion, six-year deal with Lambda — a cloud provider backed by Nvidia — for approximately 350 megawatts of capacity at a facility in Nueces County, Texas, to be built by Hut 8, a former Bitcoin miner turned data center operator.
Earlier in 2026, Broadcom disclosed an agreement to supply approximately 3.5 gigawatts of computing capacity through Google Cloud's next-generation tensor processing units, starting in 2027. Anthropic separately committed to $30 billion of Microsoft Azure compute capacity — powered by Nvidia GPUs — under a November 2025 partnership.
The $180 billion figure appears to capture these commitments in aggregate. The commitments span hyperscalers (AWS, Azure, Google Cloud) whose own chip choices include Trainium2 and TPU alongside Nvidia silicon. The 2.5 GW Nvidia-specific capacity figure is therefore the cleaner signal for Nvidia's direct revenue exposure from the Anthropic relationship.
**What "Neocloud" Means: Why It Matters Here**
Several of Anthropic's commitments run through a category of provider that did not exist at scale three years ago: the neocloud. A neocloud is a cloud operator built specifically for GPU-dense AI workloads, offering raw compute access rather than the broader managed services portfolio that hyperscalers like AWS, Azure, and Google Cloud provide. Neoclouds typically offer the same hardware at 40–70% lower per-GPU-hour pricing than hyperscalers, and they provision capacity more quickly. Nscale, Lambda, and IREN — all of which appear in Anthropic's infrastructure stack — fit this model.
Nvidia has formalized its relationship with the neocloud tier through its DSX AI Factory program, in which local operators provide land, power, and data center shell capacity while Nvidia supplies its full-stack platform: accelerated computing, networking, software, and reference designs. The Australian version of this program, announced September 9, 2026, involves eight local operators: Firmus, Sharon AI, IREN, Megaport, ResetData, CDC, NEXTDC, and AirTrunk. Together they are targeting up to 2 gigawatts of capacity by 2027 — roughly doubling Australia's existing data center footprint of approximately 1.6 GW.
Sharon AI plans to deploy up to 68,000 Nvidia GPUs using Quantum InfiniBand and Spectrum-X Ethernet networking. IREN, the Australian-born neocloud operator that pivoted from Bitcoin mining into AI infrastructure, is applying the DSX architecture to its planned 800-megawatt Bundey campus in South Australia. IREN has a deeper relationship with Nvidia than its DSX participation suggests: the two companies announced in May 2026 a $3.4 billion managed services deal under which IREN provides Nvidia with GPU compute for Nvidia's own internal AI and research workloads, plus a warrant allowing Nvidia to purchase up to 30 million IREN shares at $70 each — a potential equity investment of up to $2.1 billion if fully exercised.
The DSX model, repeated globally — with similar announcements across the Middle East, Europe, and North America — is how Nvidia's sovereign AI thesis translates into revenue: nations and regions that want domestic AI infrastructure, rather than dependence on US hyperscaler availability, become customers through local DSX operators who are themselves Nvidia hardware buyers.